GANsGANs

In 2014, a graduate student named Ian Goodfellow proposed a strikingly simple recipe for teaching machines to create. His paper, 'Generative Adversarial Nets,' pitted two neural networks against each other in a contest that improved them both.

One network, the generator, tries to produce fake data, such as images of faces, that look real. The other, the discriminator, tries to tell real examples from fakes. As the discriminator gets better at catching forgeries, the generator is forced to produce ever more convincing ones.

The story goes that Goodfellow sketched the idea after a discussion with friends at a bar, then coded a working version that same night. The adversarial setup solved a problem that had long stumped researchers: how to train a generative model without spelling out exactly what a realistic image should look like.

GANs kicked off an explosion of AI-generated imagery, from photorealistic faces of people who never existed to style transfer and image editing. They also introduced the world to convincing deepfakes, raising early alarms about synthetic media.

Training GANs proved notoriously finicky, prone to collapse and instability. In time, diffusion models would overtake them for many image tasks, but GANs had proved something profound: a machine could learn to imagine.